Commentary warned of shaky economics and rising safeguards, but our tracker shows 4,124 open AI roles across 174 employers, with no surge in safety hiring despite watermarking momentum.
A split-screen week: safeguards scale up while economics get questioned
Zvi Mowshowitz argues watermarking is here and cheap to deploy, Sebastian Raschka documents Anthropic’s implementation details, while Azeem Azhar highlights failures of multi-agent consensus. At the same time, Gary Marcus questions the revenue narratives and the political viability of data center buildouts. Practitioners like Simon Willison, Linus Torvalds, Matt Webb, and Ethan Mollick show AI quietly embedding in developer workflows and everyday chores.
Across these claims, our tracker points to a steady labor market for AI. Employers are hiring widely, particularly for modeling and engineering, with modest activity in evaluation and safety. Where the commentary foresees slowdowns or costly pivots, job openings do not yet show a retrenchment.
Watermarking is rolling out. Hiring for it is modest.
Mowshowitz’s case for watermarking is pragmatic. “AI outputs are not deterministic.” He describes a keyed pseudo-randomness scheme and notes, “You provide an API that lets anyone check for the watermark.” He emphasizes cost and fidelity: “The marginal cost of doing this is very close to zero.” He also says the governance context is settled: “The European Union Code of Practice, signed by the major Western AI labs, requires future AI models to use such watermarks.” And he points to live deployments: “Google implemented this, including for Gemini 3.7 Flash, and they have been rolling out this feature since 2024.”
Sebastian Raschka documents the mechanics on the Anthropic side, opening with, “Hi everyone. So, a few days ago, Anthropic announced that they will watermark the text outputs of their Claude models.” His piece is an explainer, not a claim about market impact.
Do our numbers show hiring moving to enable this watermarking wave? Evaluation and safety roles are present but niche in volume: 41 open, with 9 opened and 2 closed in 30 days, across 17 employers. That is small compared to core build-out. Modeling and engineering stands at 1,617 open, with 379 opened and 106 closed in 30 days, across 143 employers. If watermarking is, as Mowshowitz puts it, of “very close to zero” marginal cost, the absence of a hiring surge in evaluation and safety fits that view. The workload looks absorbed mostly by existing modeling and engineering teams rather than a large new category of roles.
Anthropic specifically has 114 open roles, with 33 opened in 30 days. Our tracker does not tag roles to watermarking, so we cannot attribute those openings to the feature. But the hiring signal is that labs linked to watermarking announcements are still adding headcount rather than freezing it.
Revenues and data centers under fire, but roles keep opening
Gary Marcus cautions against confusing revenue yardsticks. “Anthropic’s boosters are rushing to celebrate their ARR.” He warns that “Annualized Run Rate, alas, means no such thing.” and adds, “When people tell you Anthropic had this or that ARR, you really need to know which meaning they are talking about. They rarely tell you. But they really mean the latter.”
In a separate post, Marcus questions the economics and the politics of infrastructure expansion: “Either way, we are talking about current revenues in the tens of billions (or low hundreds of billions if you are really optimistic), against Capex in the trillions.” His bottom line is blunt: “You don’t have to be Einstein to see the math ain’t mathing.” He then argues momentum is turning: “Republicans are abandoning data centers like rats abandoning sinking ships.” and “Nobody wants this, and the economics make no sense.” which leads to, “Perhaps no industry in history has spoiled its own prospects faster than the AI industry.”
Our hiring data does not yet reflect that kind of pullback. Overall, our tracker counts 4,124 open AI roles across 174 employers. Infrastructure roles total 389 open, with 64 opened and 18 closed in 30 days, across 90 employers. Openings exceed closures in infrastructure, which is inconsistent with a broad industry retreat from building. Across role families, recent activity also leans expansionary: modeling and engineering opened 379 and closed 106 in the last 30 days; data opened 181 and closed 19; research opened 32 and closed 4; product and design opened 20 and closed 2; evaluation and safety opened 9 and closed 2.
Looking at specific employers, large builders and adopters are posting at scale. Accenture has 606 open roles, with 209 opened in 30 days. Capital One has 176 open, with 28 opened. OpenAI lists 148 open, with 62 opened. Anthropic is at 114 open, with 33 opened. None of these figures suggest a hiring freeze that would mirror the severe headwinds Marcus describes. That does not invalidate concerns about accounting definitions or long-horizon capex, but current employer job feeds show continued investment in teams.
One agent beats four. Our data cannot discriminate here.
Azeem Azhar highlights the hidden-profile problem in multi-agent setups: “After discussion, most model families chose correctly in only 17-36% of runs, while a single agent handed the entire evidence base got it right nearly every time. Only one model (somewhat) escaped: Mythos 5, at about 85% (why, we don’t know).” He attributes some of this to homogeneity: “First, LLMs lack diversity (they are low-variance): set 30 agents the same coding task and 18 of them will name their git branch identically.”
Our tracker does not classify roles by architectural preference such as multi-agent orchestration versus single-agent deployments. What we can say is that core delivery teams are where the hiring is: 1,617 modeling and engineering roles are open across 143 employers, and 799 data roles are open across 118 employers. If firms are simplifying to single-agent patterns, that may reduce some complexity in product and platform roles, but our data does not isolate this effect.
Coding agents and everyday help: usage is up, and teams are hiring
The practitioner stories point to real gains in productivity and learning. Simon Willison quotes Linus Torvalds’s experience: “And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work.” Torvalds also concedes limits, but credits the tool overall: “So credit where credit is due and I let the AI write the commit message above.” Matt Webb says that with ChatGPT, “So I sat down with ChatGPT and I didn’t get it to write the code, but I got it to educate me.” and concludes, “It pushes me to learn more. I like that as an outcome.” Thomas Ptacek, quoted by Willison, pushes builders to ship interfaces now that agents reduce friction: “Go build a native UI. It’ll probably change the way you think.”
Ethan Mollick points to unglamorous, useful automation: “In terms of everyday usefulness and saving time, Codex & Claude Code are very capable of doing the thing where you ask them to "fill out the forms that I got an email about" and they do it well & without further intervention.” He adds, “Really nice for low-risk time-consuming stuff that scattered attention.” and predicts sentiment-use gaps: “Its going to be an era of contradictions. Polls will show everyone hates AI overall but also everyone will secretly use AI all the time for lots of stuff.” He extends that to consumer support: “Life is full of things that, by complexity or design or lack of care or required time, are hard to navigate: healthcare, government, personal finance, school forms are all among them It is why I feel consumer AI is underrated. People muddle by, but are missing support & AI is good enough to help.”
Do employers seem to be staffing to capture these gains? The bulk of openings are still in core build functions. Modeling and engineering has 1,617 open roles; data has 799; research has 223. Product and design shows 98 open roles, with 20 opened and 2 closed in 30 days, across 49 employers. That is smaller than engineering, which is consistent with a phase where firms are still wiring agents into systems and data, even as UI experimentation becomes cheaper. Services and adopters are well represented among top posters, with Accenture at 606 open and PwC at 91. Capital One has 176, and Databricks has 89. These employers are positioned to translate the “debug session from hell” and “fill out the forms” gains into standardized offerings, and they are hiring to do so.
What holds up and what does not
- Watermarking: Mowshowitz’s claim that it is low-cost to implement aligns with our hiring mix. We do not see a distinct surge in evaluation and safety hiring. Evaluation and safety have 41 open roles today, versus 1,617 in modeling and engineering, suggesting the work is being absorbed by core teams.
- Revenue skepticism and infrastructure politics: Marcus’s warnings are not mirrored in the job feeds we track. Infrastructure openings outpace closures, and leading labs and adopters continue to post new roles.
- Agent design choices: Azhar’s results are notable, but our dataset cannot validate whether teams are consolidating around single-agent approaches.
- Everyday and developer use: The Torvalds, Webb, Ptacek, and Mollick accounts match ongoing demand for engineers and data talent. Employers are still adding those roles, which is consistent with making these gains repeatable at scale.
Azeem Azhar’s broader critique in a separate post targets the industry narrative itself: “They are the titans of AI, the bosses of the labs, the investors behind them.” who needed to build, “To do this, they would need capital: to write their software and to build 21st-century infrastructure to run it.” and remind us “But they warned that this was no ordinary software.” That tension between promises and delivery is real. The hiring market, however, shows continued commitment to delivery. Until employer job feeds reverse, the labor signal supports the view that firms are still building, even as the story around them gets rougher.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: AI Text Watermarking Is Free And Good
- Sebastian Raschka: How Claude Watermarks AI-Generated Text
- Azeem Azhar: 🔮 Why one AI is better than four #598, 🏦 The problem with petards
- Simon Willison: llm 0.33, Quoting Linus Torvalds, Quoting Matt Webb, Stop Making TUIs, More than just code review
- Gary Marcus: ARR vs ARR. Watch out for this one sly trick., Data center madness
- Ethan Mollick: In terms of everyday usefulness and saving time, Codex & Claude Code a, [Its going to be an era of contradictions.
Polls will show everyone ha](https://bsky.app/profile/emollick.bsky.social/post/3mtms457ksk2c), Life is full of things that, by complexity or design or lack of care o